The challenge for classification methods often lies in constructing the classifier based on the probability distribution of the sample data, which can lead to significant errors, especially with small or high-dimensional sample sets. In this work, we employed a modified hybrid Radial Basis Function-BackPropagation (RBF-BP) classifier method called the mHybrid classifier to solve the neural network’s nonlinear classification problems. The mHybrid classifier serves as a density cluster-generating approach, ddetermining the number of RBF hidden neurons while maintaining the self-learning process and updating the neural network’s weights through BP. In the final step of the BP neural network in the mHybrid classifier, the centers and width parameters of the basis functions are iteratively updated using the stochastic gradient descent algorithm. Additionally, we have utilized an optimized sample filling and screening method to generate synthetic samples after the RBF process. This approach, called the Sample Generation (SG) method, significantly improves classification performance in sparse datasets. Our experiments on artificial datasets and several UCI benchmark datasets which combine the mHybrid classifier and the SG method have demonstrated the effectiveness of our implementation strategies. The experimental results also showed that the implemented algorithm improved accuracy and has broader applicability.